Glossary term
Glossary term
Safety and Alignment
A controlled capability to disable, isolate, roll back, or restrict an AI system when risk exceeds acceptable thresholds or when an incident requires rapid containment. A kill switch is only useful if ownership, authority, technical implementation, rollback steps, and business continuity impacts are tested before an incident.
Microsoft and Google publicly committed at the 2023 White House AI Voluntary Commitments to implement and exercise emergency response capabilities including kill switch functionality.
The Seoul AI Safety Summit (May 2024) Frontier AI Safety Commitments included pause and emergency disabling commitments from 16 leading developers.
Tessa, the National Eating Disorders Association chatbot, was disabled by NEDA in May 2023 within days of giving harmful weight-loss advice.